Author(s): Yifan Yang

Linked Author(s):

Keywords: Convolutional autoencoder Flow-field reconstruction Physics-informed training Coarse-grain bed

Abstract: This study introduces a convolutional autoencoder-based neural network model for reconstructing real-time velocity and turbulent kinetic energy (TKE) fields over coarse-grain beds using boundary profiles. The overall aim is to establish an efficient and extensible model architecture for handling various input/output data formats for river flows. The customized physics-informed training strategy enhances the training efficiency and improves accuracy. Comprehensive assessments demonstrate that the model has satisfactory accuracy in capturing spatiotemporal flow patterns, particularly for horizontal and total velocities (e. g. wake zone). Time-averaging using instantaneous fields can further improve the model's spatial accuracy, bringing the results closer to target fields as outputs accumulate. The model also shows limitations in capturing the scattering patterns of vertical velocities and high-TKE regions. Robustness analyses reveal that the velocity-trained model can generally maintain stable performance when subjected to incomplete or corrupted inputs. The fidelity of reconstructed TKE fields is more sensitive to those disturbances than velocity fields, which necessitate cleaner inputs and enhanced training. Generally, the proposed model offers a promising approach to real-time flow field reconstruction that is comparable in accuracy to traditional CFD models.

DOI:

Year: 2025